迷失于压缩：对抽取式提示压缩器的受控跨语言审计

摘要

抽取式提示压缩有望通过去除低信息量词元来降低大语言模型的推理成本，而诸如LLMLingua-2等学习型压缩器在英语基准上报告了强劲的性能。大多数其他语言已经支付了词元溢价：相同内容的词元消耗比英语多1.3至1.8倍。我们探究的问题是：压缩是缩小还是扩大了这一差距。我们使用涵盖五种文字的十种语言的完全平行数据，在目标模型分词器中进行了预算匹配控制，将四种学习型压缩器与四种确定性基线进行对比审计，涉及来自十家供应商的十一个目标模型（超过25万次评估调用）。其中三种压缩器使用英语监督训练（LLMLingua-2 XLM-R/mBERT；来自生产级Headroom技术栈的Kompress-v2）；第四种XProvence则采用多语言训练。首先，迁移差距真实存在，跨目标模型和压缩器骨干网络均可复现，且强烈依赖于保留率：在0.33的保留率下，英语保留了57-62%的归一化上下文利用率，而立陶宛语仅保留10-24%，中文则几乎为零——尽管中文的词元溢价最小。其次，这一差距与压缩监督数据相关，而非架构。所有三种英语训练的压缩器均表现出该差距，确定性方法未出现可比差距，而多语言训练的XProvence v1则完全没有差距。其v2版本在翻译数据上重新训练后，在激进阈值下清空了92%的中文上下文，且毫无预警。第三，在难度更高的长上下文场景中，激进的学习型压缩导致五种非英语语言中的三种其压缩后上下文效用降至与无上下文相当或更低。先翻译后压缩的流水线在五种测试语言中的三种中，以约一半的词元成本达到或超越了原生压缩的效果。我们公开了所有代码、压缩结果和模型输出。英语之外的安全压缩预算要小得多。

English

Extractive prompt compression promises to cut LLM inference costs by removing low-information tokens, and learned compressors such as LLMLingua-2 report strong results on English benchmarks. Most other languages already pay a token premium: the same content costs 1.3-1.8x more tokens than in English. We ask whether compression closes or widens this gap. Using fully parallel data in ten languages spanning five scripts, with controls budget-matched in the target model's tokenizer, we audit four learned compressors against four deterministic baselines, on eleven target models from ten vendors (over 250,000 evaluation calls). Three of the compressors are trained with English supervision (LLMLingua-2 XLM-R/mBERT; Kompress-v2 from the production Headroom stack); the fourth, XProvence, is trained multilingually. First, the transfer gap is real, replicates across target models and compressor backbones, and is strongly rate-dependent: at a 0.33 keep-rate English retains 57-62% of normalized context utilization while Lithuanian retains 10-24% and Chinese essentially none, despite Chinese having the smallest token premium. Second, the gap tracks compression supervision data, not architecture. All three English-trained compressors show it, deterministic methods show no comparable gap, and the multilingually trained XProvence v1 shows none. Its v2 release, retrained on translated data, empties 92% of Chinese contexts at its aggressive threshold without any warning. Third, in a harder long-context setting, aggressive learned compression drives compressed contexts to or below no-context utility in three of five non-English languages. A translate-then-compress pipeline matches or beats native compression at roughly half the token cost in three of five tested languages. We release all code, compressions, and model outputs. Safe compression budgets are much smaller outside English.